Credit Risk Assessment Model Using Machine Learning

Abstract

An individual's creditworthiness is quantified by their credit score, which is based on their credit history. Credit ratings are used by financial companies to distinguish between debtors who will fulfill their obligations and those who won't. This system has not yet been digitized or used in Kenya's banking sector. The study's objective is to develop a reliable credit scoring model that will give organizations like these a reliable reference score to rely on when verifying a client. The data used included customer details including age, loan amount, marital status, and sex, among other factors. It was collected from Kenyan commercial banks between 2016 and 2021. It was possible to create an optimized model with an accuracy of 93%. The model was built using the Gradient Boosting method (GBM) and is based on classification and Regression Trees (CART). In addition, a new hybrid model with a two-step architecture is proposed. The first uses distributed random forests, where each decision tree's output is fed into a deep neural network (DNN) that has been trained to outperform the random forest method on its own. Giving a rating without adequate rationale is unethical because a person's creditworthiness is a sensitive matter. An examination of the model's interpretability was conducted, and the results created visual representations of the factors that influence the model's output and the data required for a successful client analysis. The outcomes were reliable and accurately replicated the process of appraising an individual. The proposed model could be put into practice in order to give failure evaluation and prediction a real-world foundation. Simultaneously provide a thorough explanation of the outcomes at the same time. This could considerably aid financial organizations in preventing the loss of millions of dollars from non-performing loans.